Collaborative filtering (CF) is widely applied in recommender systems to predict user preferences or interests according to a user’s historical information. Traditional CF methods mainly adopt batch-processing methods to train recommender models, which require prior preparation of all training data. However, user preferences or interests always change with time, and it is impossible to prepare all of the training data at once. In actuality, the data are obtained with time with a certain sequence. Therefore, to update the model, the trained model needs to be re-trained according to all of the historical datasets. As a result, the cost of re-training is very high, which slows down the recommender model, and makes new updates of the user data difficult to capture. To solve these problems, the online recommender system emerged. In this paper, we propose a confidence-weighted bias model (CWBM) for online collaborative filtering (OCF). This model adds bias into CF and further introduces confidence weights; thus, it can improve the stability and accuracy of OCF. A comparative experiment on two real datasets, Movielens100K and Mvielens1M, show that the method proposed in this paper is superior to other baseline methods.